Yup, that’s exactly what this is. If you’ve been using generative models since the early Stable Diffusion days, it’s a pretty common (and useful!) technique: using a sketch (SVG, drawn, etc) as an ad-hoc "controlnet" to guide the generative model’s output.
Example: In the past I'd use a similar approach to lay out architectural visualizations. If you wanted a couch, chair, or other furniture in a very specific location, you could use a tool like Poser to build a simple scene as an approximation of where you wanted the major "set pieces". From there, you could generate a depth map and feed that into the generative model, at the time SDXL, to guide where objects should be placed.
I’ve always said that the more divergent the input is from the resulting output, then the less personal expression you have. For me, in order of moving away from meaningful control in generative models, it goes: “text → code,” “text → picture,” and, at the very bottom, “text → music.”
For me personally, music composition begins and ends with the motif - the melody itself. It’s the part I enjoy the most, and it’s also the part I have the most individual control over since I can sing.
Everybody makes music differently, but if you lack the ability to play an instrument and you also can’t whistle or sing, it’s hard for me to imagine how you’d have any meaningful control over the melody.
How would a non‑musician express an actual melody that they came up with (beyond simple things like instrumentation and general “feelings”) in text? RED RED RED BLUE. (Sorry couldn't resist a Mission Hill reference here.)
With all that out of the way, there's still lots of room for using AI in music. I’ve used it to take some of my existing songs, mostly pianistic in nature, and swap out instrumentation and arrangements just to play around with different soundscapes. It's like BIAB on steroids.
Rosetta Stone's code, by virtue of being a pretty popular community effort, is also very well-vetted. Even when I've tested some of the examples in more obscure langs like Malbolge, they've always worked.
Same, I really like the ONNX format. I only wish that they weren't so frustratingly difficult to use on Apple iOS. Their browser engine, WebKit, has become annoyingly restrictive over the years in terms of the working memory cap.
I ran into quite a few out-of-memory iOS safari issues when I was building continuous voice recognition for my blind chess game, so people could play while on the go.
So I use a VAD onnx (Silero [1]) to automatically detect when someone is talking, and then it sends the audio into one of the voice recognition libraries.
I originally tried to get away with just Whisper Tiny in the chess game [2], but it performs worse on the kinds of short phrases (knight E4, c takes d5, etc) used to dictate chess notation. Even with hotword-based phrasing and corrections, I found its accuracy on brief inputs noticeably poorer. So I switched over to Sherpa [3] trained on gigaspeech. It’s significantly more accurate, but it also comes with a correspondingly larger memory footprint.
Ideally, I would have used just one engine, but I needed a fallback for iOS devices (especially older ones) which can easily OOM.
Neat. As mentioned in the article, it kind of reminds me of Zombies, Run!, a mobile fitness game from 2012 that was very audio-immersive. It would make it sound like the zombies were getting closer to you, so you’d be motivated to run faster.
From the article:
> The narrative_director node builds a structured prompt from the session state and calls Gemini 2.5 Flash with temperature=1.2 and max_output_tokens=200. High temperature because we want genuine variety.
I'd be genuinely curious to see a couple of the stories. I looked at the seeds they're using, and they're pretty open-ended. In my experience, though, the less specific and more generalized your initial prompt (especially for storytelling) the more tapiocally generic the output tends to be, even when the temperature is dialed way up.
On a slightly related note, I actually pre-sort my electronic music playlists into rough BPM ranges in intervals of 10 (e.g. 120, 130, 140, etc.) Depending on the pace I want to set, I can then pick a playlist where it’s easy to time my footfalls exactly to the bass pulse. Makes sprint/runs more entertaining for a hardcore Pump It Up fanatic.
I did briefly play around with the idea of using a BLE heart-rate monitor to autopick music according to my given heart rate, but there'd be a lot of mental whiplash when doing HIIT (high-intensity interval training).
This feels icky, and it kind of goes against Show HN guidelines which is meant for things you actually created. It looks like you're taking the source code for various games made by other people and are hosting them.
But this post gives the impression that you made the game.
You've also posted 3 separate Show HNs in the span of 24 hours.
I don’t, personally. I’ll sometimes attach articles I’ve written if they’re apropos to a specific discussion (usu. in the context of a longer reply), but I prefer to let that happen organically rather than force it.
So I see pretty modest traffic to my site. Discounting bots/scrapers, it's about a thousand visits a month, so it's not a lot. I also don't have a newsletter, which definitely doesn't help.
Honestly, you really should focus on a niche. For example, Simon Willison [1] keeps his blog super consistent, he's the AI journalist. Josh Comeau [2] is another good example from Hacker News: probably 99% of his content is front-end design and development. Becoming (or establishing oneself) as an expert in some space goes a long way toward increasing your subscriber count.
My articles are mostly satirical in nature and have all determinism of a literary brownian motion machine (writing on algorithms, thought experiments, fiction, etc.) so there’s not a particularly cohesive identity to the site.
Another thing that is often mentioned is that you need to publish regularly and often. I tend to be a perfectionist, and spend a lot of time crafting each article with custom illustrations, animations, and interactive elements. This works for people like Bartosz Ciechanowski [3], but I'm not sure it helps drive traffic for me personally.
Nice job. I'd seen similar things in the past but I had no idea this was an actual type of puzzle.
Your game has a bug which is relatively common in drag-and-drop games. If the viewport is smaller and I drag a match off-screen by accident, it can get stuck to the edge. Then it can become detached in such a way that makes it impossible to recover the match.
Thanks for trying this out. I'll fix the off-screen issue later.
I created this game because of an old news my wife shared with me, it was about The New York Times bought the Wordle game. I was then surprised that such a simple, old-style game caught the attention of a big company. I said "I'd rather play a matchstick puzzle", and it was that moment that inspired me to create a free online version with all the puzzles procedurally generated by my own hand-made solver (I like the sound of this :D). Personally, I like moving one and only one matchstick to solve a puzzle rather than 2 or 3 sticks, also I don't like those tricky solutions like negative number, rotation trick and etc. although they could be a very creative solution and let you think outside the box sometimes, IMO, it is just not elegant and sometime distracting to encourage people to think in a systematic way.
You made me chuckle a bit with the whole “hand-made solver” thing with discrete puzzles like this, since 99% of the time to create a new puzzle, you just start with a solved problem (arithmetic equality) and introduce a sequence of random valid moves to put it in a unsolved state.
I remember one of my freshman computer science assignments was to make a Rubik’s Cube game. That was when a lot of undergrads were first introduced to the idea of using a stack and then "unwinding" it to simultaneously build the problem + solution.
yeah, the solver behind Mathstick is pretty much like what you described: building the problem + solution simultaneously, besides you also need to collect the multiple solutions for the same question and information for further classifying the difficulty of a puzzle.
Just like "hand-made" pizza used as a response to the rise of frozen, factory-produced pizza and to establish authenticity against fast-food chains. Maybe "hand-made" will be used to emphasize a app/software is made by human rather than being vibecoded or completely AI generated in the future. Also, if you use the Puzzle Maker to create your own puzzle, it becomes a "hand-made" puzzle :)
Nice article. Chess roguelike is an incredibly fun genre, and there are several good games in this space.
It’s also pretty easy to come up with all kinds of fun variants for them, such as:
- Kings are strong, charismatic leaders and can persuade enemy units to join their side.
- Rooks are defensive units and can boost the defenses of adjacent pieces, or they’re strong enough to shift allies by literally pushing them in a direction as part of their turn.
- Bishops are men of the cloth and can heal the wounded. Pieces that are resurrected are vulnerable to being “turned” by bishops.
The difficult part is playtesting the everloving sh## out of it, since it’s a bit like coming up with new mechanics or cards for MTG: easy to do, but hard to balance.
Thank you. I agree. When done well, a good mashup can feel “easy” and “obvious” but really the devil is in the details. Creating the perfect balance of features across the whole design space for a shippable game takes extreme skill, taste and a lot of effort.
Garry often describes the fast iterative development loop for startups. Launch, test and get user feedback, polish off the rough edges and then do it again. I believe the same process is necessary for excellent game development.
> Show HN is for something you've made that other people can play with. Off topic: blog posts, sign-up pages, newsletters, lists, and other reading material. Those can't be tried out, so can't be Show HNs. Make a regular submission instead.
Example: In the past I'd use a similar approach to lay out architectural visualizations. If you wanted a couch, chair, or other furniture in a very specific location, you could use a tool like Poser to build a simple scene as an approximation of where you wanted the major "set pieces". From there, you could generate a depth map and feed that into the generative model, at the time SDXL, to guide where objects should be placed.
reply